Atmosphere Ocean Science Colloquium
Combining physical and machine learning forecasts for Earth system prediction
Speaker: Eviatar Bach, CalTech
Location: 60 Fifth Avenue 7th Floor Open Space
Date: Wednesday, March 29, 2023, 2:30 p.m.
Synopsis:
Machine learning (ML) holds the potential to improve Earth system prediction by
learning directly from data, bypassing deficiencies in existing dynamical models. Hybrid
methods, which combine ML with dynamical models, leverage the strengths of both approaches.
I will present two such hybrid methods that use tools from data assimilation: Ensemble
Oscillation Correction (EnOC) and the Multi-Model Ensemble Kalman Filter (MM-EnKF).
Oscillatory modes in the climate system are an important source of predictability beyond the
weather timescale, and these modes can often be predicted by data-driven methods with higher
skill than dynamical models. However, they only represent a portion of the variance of the
signal, and a method for beneficially combining them with dynamical forecasts of the full system
has not previously been developed. Ensemble Oscillation Correction (EnOC: Bach et al., 2021)
is a hybrid forecasting method for combining ML forecasts of specific modes with a full-field
dynamical model. I will show results of EnOC applied to forecasts of South Asian monsoon
rainfall, significantly outperforming the state-of-the-art forecasts on subseasonal-to-seasonal
timescales.
A more general method for integrating dynamical models, ML forecasts, and observations is
multi-model data assimilation (MM-DA). MM-DA generalizes the variational, Bayesian, and
minimum variance formulation of the Kalman filter. I will show how multiple model ensembles
can be combined for both DA and forecasting in a flow-dependent manner using the Multi-
Model Ensemble Kalman filter (MM-EnKF: Bach and Ghil, 2023). In numerical experiments
with multiscale chaotic models, the MM-EnKF is shown to improve probabilistic predictions
using multiple models with parametric error, different resolved scales, and different fidelities.
I will conclude with a discussion of the future prospects for hybrid forecasting of the Earth
system.